Cross-Motor Generalization of Data-Driven Fault Diagnosis in Multi-Motor Manufacturing Conveyor Systems

This study evaluated single-source cross-motor transfer behaviour in data-driven fault diagnosis using manufacturing-conveyor sensor records. Motor 3 (M3), the only motor with both normal-operation and recorded stall data, was used as the source domain. The source dataset comprised 317 one-minute normal-operation windows and 178 one-minute recorded stall windows. Target stall datasets comprised 181 windows for Motor 1 (M1), 71 for Motor 2 (M2), and 77 for Motor 4 (M4). Temperature, the recorded Raw channel, and x, y, and z directional measurements were retained as predictor variables; a resultant magnitude was derived for descriptive analysis. Initialization records were excluded, and the remaining observations were aggregated into one-minute median windows. Six classifiers—Logistic Regression, k-nearest neighbours, Decision Tree, Random Forest, Gradient Boosting, and support vector machine—were evaluated on a class-preserving chronological 70:30 M3 source split. All six models achieved 1.000 accuracy, precision, recall, F1-score, and ROC-AUC on held-out source windows. Transfer to target stall events revealed model- and motor-dependent differences. M1 and M4 were classified as stall by all six models, whereas M2 produced complete disagreement across some algorithms: Logistic Regression and KNN classified 0% of M2 windows as stall, while the other four classifiers classified 100%. Four-motor Kruskal–Wallis analyses showed substantial motor-level distribution differences, particularly for x (ε²=0.989) and y (ε²=0.995). These findings show that perfect validation on one source motor does not establish generalization to other motors. Because target normal-operation records were unavailable and the analysis used one source motor, the results constitute a single-source transfer baseline rather than validated multi-motor diagnostic performance.